token limiting

diff --git a/src/paperless_ai/ai_classifier.py b/src/paperless_ai/ai_classifier.py
index 26c0f18731595afbeb2174af630bd1ebbfc39e09..fd660763c020e36d8770e11bf99a4c0bc7d434ca 100644 (file)
--- a/src/paperless_ai/ai_classifier.py
+++ b/src/paperless_ai/ai_classifier.py
@@ -9,13 +9,14 @@ from documents.permissions import get_objects_for_user_owner_aware
 from paperless.config import AIConfig
 from paperless_ai.client import AIClient
 from paperless_ai.indexing import query_similar_documents
+from paperless_ai.indexing import truncate_content
 
 logger = logging.getLogger("paperless_ai.rag_classifier")
 
 
 def build_prompt_without_rag(document: Document) -> str:
     filename = document.filename or ""
-    content = document.content or ""
+    content = truncate_content(document.content or "")
 
     prompt = f"""
     You are an assistant that extracts structured information from documents.
@@ -48,20 +49,20 @@ def build_prompt_without_rag(document: Document) -> str:
     {filename}
 
     CONTENT:
-    {content[:8000]}
+    {content}
     """
 
     return prompt
 
 
 def build_prompt_with_rag(document: Document, user: User | None = None) -> str:
-    context = get_context_for_document(document, user)
+    context = truncate_content(get_context_for_document(document, user))
     prompt = build_prompt_without_rag(document)
 
     prompt += f"""
 
     CONTEXT FROM SIMILAR DOCUMENTS:
-    {context[:4000]}
+    {context}
     """
 
     return prompt

diff --git a/src/paperless_ai/indexing.py b/src/paperless_ai/indexing.py
index 4e0f2ffdf58e3a5f0fbcfbc48ce2cc63d8f46c9b..3afa1e990f49bdb66b09b5031c753c2bc4740c15 100644 (file)
--- a/src/paperless_ai/indexing.py
+++ b/src/paperless_ai/indexing.py
@@ -10,11 +10,14 @@ from llama_index.core import Document as LlamaDocument
 from llama_index.core import StorageContext
 from llama_index.core import VectorStoreIndex
 from llama_index.core import load_index_from_storage
+from llama_index.core.indices.prompt_helper import PromptHelper
 from llama_index.core.node_parser import SimpleNodeParser
+from llama_index.core.prompts import PromptTemplate
 from llama_index.core.retrievers import VectorIndexRetriever
 from llama_index.core.schema import BaseNode
 from llama_index.core.storage.docstore import SimpleDocumentStore
 from llama_index.core.storage.index_store import SimpleIndexStore
+from llama_index.core.text_splitter import TokenTextSplitter
 from llama_index.vector_stores.faiss import FaissVectorStore
 
 from documents.models import Document
@@ -220,6 +223,23 @@ def llm_index_remove_document(document: Document):
     index.storage_context.persist(persist_dir=settings.LLM_INDEX_DIR)
 
 
+def truncate_content(content: str) -> str:
+    prompt_helper = PromptHelper(
+        context_window=8192,
+        num_output=512,
+        chunk_overlap_ratio=0.1,
+        chunk_size_limit=None,
+    )
+    splitter = TokenTextSplitter(separator=" ", chunk_size=512, chunk_overlap=50)
+    content_chunks = splitter.split_text(content)
+    truncated_chunks = prompt_helper.truncate(
+        prompt=PromptTemplate(template="{content}"),
+        text_chunks=content_chunks,
+        padding=5,
+    )
+    return " ".join(truncated_chunks)
+
+
 def query_similar_documents(
     document: Document,
     top_k: int = 5,
@@ -247,7 +267,9 @@ def query_similar_documents(
         doc_ids=doc_node_ids,
     )
 
-    query_text = (document.title or "") + "\n" + (document.content or "")
+    query_text = truncate_content(
+        (document.title or "") + "\n" + (document.content or ""),
+    )
     results = retriever.retrieve(query_text)
 
     document_ids = [